1 /*
2 * Copyright (c) 2015 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #include <stdlib.h>
25 #include <math.h>
26
27 #include "util/macros.h"
28 #include "main/macros.h"
29
30 #include "gen_l3_config.h"
31
32 /**
33 * IVB/HSW validated L3 configurations. The first entry will be used as
34 * default by gen7_restore_default_l3_config(), otherwise the ordering is
35 * unimportant.
36 */
37 static const struct gen_l3_config ivb_l3_configs[] = {
38 /* SLM URB ALL DC RO IS C T */
39 {{ 0, 32, 0, 0, 32, 0, 0, 0 }},
40 {{ 0, 32, 0, 16, 16, 0, 0, 0 }},
41 {{ 0, 32, 0, 4, 0, 8, 4, 16 }},
42 {{ 0, 28, 0, 8, 0, 8, 4, 16 }},
43 {{ 0, 28, 0, 16, 0, 8, 4, 8 }},
44 {{ 0, 28, 0, 8, 0, 16, 4, 8 }},
45 {{ 0, 28, 0, 0, 0, 16, 4, 16 }},
46 {{ 0, 32, 0, 0, 0, 16, 0, 16 }},
47 {{ 0, 28, 0, 4, 32, 0, 0, 0 }},
48 {{ 16, 16, 0, 16, 16, 0, 0, 0 }},
49 {{ 16, 16, 0, 8, 0, 8, 8, 8 }},
50 {{ 16, 16, 0, 4, 0, 8, 4, 16 }},
51 {{ 16, 16, 0, 4, 0, 16, 4, 8 }},
52 {{ 16, 16, 0, 0, 32, 0, 0, 0 }},
53 {{ 0 }}
54 };
55
56 /**
57 * VLV validated L3 configurations. \sa ivb_l3_configs.
58 */
59 static const struct gen_l3_config vlv_l3_configs[] = {
60 /* SLM URB ALL DC RO IS C T */
61 {{ 0, 64, 0, 0, 32, 0, 0, 0 }},
62 {{ 0, 80, 0, 0, 16, 0, 0, 0 }},
63 {{ 0, 80, 0, 8, 8, 0, 0, 0 }},
64 {{ 0, 64, 0, 16, 16, 0, 0, 0 }},
65 {{ 0, 60, 0, 4, 32, 0, 0, 0 }},
66 {{ 32, 32, 0, 16, 16, 0, 0, 0 }},
67 {{ 32, 40, 0, 8, 16, 0, 0, 0 }},
68 {{ 32, 40, 0, 16, 8, 0, 0, 0 }},
69 {{ 0 }}
70 };
71
72 /**
73 * BDW validated L3 configurations. \sa ivb_l3_configs.
74 */
75 static const struct gen_l3_config bdw_l3_configs[] = {
76 /* SLM URB ALL DC RO IS C T */
77 {{ 0, 48, 48, 0, 0, 0, 0, 0 }},
78 {{ 0, 48, 0, 16, 32, 0, 0, 0 }},
79 {{ 0, 32, 0, 16, 48, 0, 0, 0 }},
80 {{ 0, 32, 0, 0, 64, 0, 0, 0 }},
81 {{ 0, 32, 64, 0, 0, 0, 0, 0 }},
82 {{ 24, 16, 48, 0, 0, 0, 0, 0 }},
83 {{ 24, 16, 0, 16, 32, 0, 0, 0 }},
84 {{ 24, 16, 0, 32, 16, 0, 0, 0 }},
85 {{ 0 }}
86 };
87
88 /**
89 * CHV/SKL validated L3 configurations. \sa ivb_l3_configs.
90 */
91 static const struct gen_l3_config chv_l3_configs[] = {
92 /* SLM URB ALL DC RO IS C T */
93 {{ 0, 48, 48, 0, 0, 0, 0, 0 }},
94 {{ 0, 48, 0, 16, 32, 0, 0, 0 }},
95 {{ 0, 32, 0, 16, 48, 0, 0, 0 }},
96 {{ 0, 32, 0, 0, 64, 0, 0, 0 }},
97 {{ 0, 32, 64, 0, 0, 0, 0, 0 }},
98 {{ 32, 16, 48, 0, 0, 0, 0, 0 }},
99 {{ 32, 16, 0, 16, 32, 0, 0, 0 }},
100 {{ 32, 16, 0, 32, 16, 0, 0, 0 }},
101 {{ 0 }}
102 };
103
104 /**
105 * Return a zero-terminated array of validated L3 configurations for the
106 * specified device.
107 */
108 static const struct gen_l3_config *
get_l3_configs(const struct gen_device_info * devinfo)109 get_l3_configs(const struct gen_device_info *devinfo)
110 {
111 switch (devinfo->gen) {
112 case 7:
113 return (devinfo->is_baytrail ? vlv_l3_configs : ivb_l3_configs);
114
115 case 8:
116 return (devinfo->is_cherryview ? chv_l3_configs : bdw_l3_configs);
117
118 case 9:
119 return chv_l3_configs;
120
121 default:
122 unreachable("Not implemented");
123 }
124 }
125
126 /**
127 * L1-normalize a vector of L3 partition weights.
128 */
129 static struct gen_l3_weights
norm_l3_weights(struct gen_l3_weights w)130 norm_l3_weights(struct gen_l3_weights w)
131 {
132 float sz = 0;
133
134 for (unsigned i = 0; i < GEN_NUM_L3P; i++)
135 sz += w.w[i];
136
137 for (unsigned i = 0; i < GEN_NUM_L3P; i++)
138 w.w[i] /= sz;
139
140 return w;
141 }
142
143 /**
144 * Get the relative partition weights of the specified L3 configuration.
145 */
146 struct gen_l3_weights
gen_get_l3_config_weights(const struct gen_l3_config * cfg)147 gen_get_l3_config_weights(const struct gen_l3_config *cfg)
148 {
149 if (cfg) {
150 struct gen_l3_weights w;
151
152 for (unsigned i = 0; i < GEN_NUM_L3P; i++)
153 w.w[i] = cfg->n[i];
154
155 return norm_l3_weights(w);
156 } else {
157 const struct gen_l3_weights w = { { 0 } };
158 return w;
159 }
160 }
161
162 /**
163 * Distance between two L3 configurations represented as vectors of weights.
164 * Usually just the L1 metric except when the two configurations are
165 * considered incompatible in which case the distance will be infinite. Note
166 * that the compatibility condition is asymmetric -- They will be considered
167 * incompatible whenever the reference configuration \p w0 requires SLM, DC,
168 * or URB but \p w1 doesn't provide it.
169 */
170 float
gen_diff_l3_weights(struct gen_l3_weights w0,struct gen_l3_weights w1)171 gen_diff_l3_weights(struct gen_l3_weights w0, struct gen_l3_weights w1)
172 {
173 if ((w0.w[GEN_L3P_SLM] && !w1.w[GEN_L3P_SLM]) ||
174 (w0.w[GEN_L3P_DC] && !w1.w[GEN_L3P_DC] && !w1.w[GEN_L3P_ALL]) ||
175 (w0.w[GEN_L3P_URB] && !w1.w[GEN_L3P_URB])) {
176 return HUGE_VALF;
177
178 } else {
179 float dw = 0;
180
181 for (unsigned i = 0; i < GEN_NUM_L3P; i++)
182 dw += fabs(w0.w[i] - w1.w[i]);
183
184 return dw;
185 }
186 }
187
188 /**
189 * Return a reasonable default L3 configuration for the specified device based
190 * on whether SLM and DC are required. In the non-SLM non-DC case the result
191 * is intended to approximately resemble the hardware defaults.
192 */
193 struct gen_l3_weights
gen_get_default_l3_weights(const struct gen_device_info * devinfo,bool needs_dc,bool needs_slm)194 gen_get_default_l3_weights(const struct gen_device_info *devinfo,
195 bool needs_dc, bool needs_slm)
196 {
197 struct gen_l3_weights w = {{ 0 }};
198
199 w.w[GEN_L3P_SLM] = needs_slm;
200 w.w[GEN_L3P_URB] = 1.0;
201
202 if (devinfo->gen >= 8) {
203 w.w[GEN_L3P_ALL] = 1.0;
204 } else {
205 w.w[GEN_L3P_DC] = needs_dc ? 0.1 : 0;
206 w.w[GEN_L3P_RO] = devinfo->is_baytrail ? 0.5 : 1.0;
207 }
208
209 return norm_l3_weights(w);
210 }
211
212 /**
213 * Get the default L3 configuration
214 */
215 const struct gen_l3_config *
gen_get_default_l3_config(const struct gen_device_info * devinfo)216 gen_get_default_l3_config(const struct gen_device_info *devinfo)
217 {
218 /* For efficiency assume that the first entry of the array matches the
219 * default configuration.
220 */
221 const struct gen_l3_config *const cfg = get_l3_configs(devinfo);
222 assert(cfg == gen_get_l3_config(devinfo,
223 gen_get_default_l3_weights(devinfo, false, false)));
224 return cfg;
225 }
226
227 /**
228 * Return the closest validated L3 configuration for the specified device and
229 * weight vector.
230 */
231 const struct gen_l3_config *
gen_get_l3_config(const struct gen_device_info * devinfo,struct gen_l3_weights w0)232 gen_get_l3_config(const struct gen_device_info *devinfo,
233 struct gen_l3_weights w0)
234 {
235 const struct gen_l3_config *const cfgs = get_l3_configs(devinfo);
236 const struct gen_l3_config *cfg_best = NULL;
237 float dw_best = HUGE_VALF;
238
239 for (const struct gen_l3_config *cfg = cfgs; cfg->n[GEN_L3P_URB]; cfg++) {
240 const float dw = gen_diff_l3_weights(w0, gen_get_l3_config_weights(cfg));
241
242 if (dw < dw_best) {
243 cfg_best = cfg;
244 dw_best = dw;
245 }
246 }
247
248 return cfg_best;
249 }
250
251 /**
252 * Return the size of an L3 way in KB.
253 */
254 static unsigned
get_l3_way_size(const struct gen_device_info * devinfo)255 get_l3_way_size(const struct gen_device_info *devinfo)
256 {
257 if (devinfo->is_baytrail)
258 return 2;
259
260 else if (devinfo->gt == 1 ||
261 devinfo->is_cherryview ||
262 devinfo->is_broxton)
263 return 4;
264
265 else
266 return 8 * devinfo->num_slices;
267 }
268
269 /**
270 * Return the unit brw_context::urb::size is expressed in, in KB. \sa
271 * gen_device_info::urb::size.
272 */
273 static unsigned
get_urb_size_scale(const struct gen_device_info * devinfo)274 get_urb_size_scale(const struct gen_device_info *devinfo)
275 {
276 return (devinfo->gen >= 8 ? devinfo->num_slices : 1);
277 }
278
279 unsigned
gen_get_l3_config_urb_size(const struct gen_device_info * devinfo,const struct gen_l3_config * cfg)280 gen_get_l3_config_urb_size(const struct gen_device_info *devinfo,
281 const struct gen_l3_config *cfg)
282 {
283 /* From the SKL "L3 Allocation and Programming" documentation:
284 *
285 * "URB is limited to 1008KB due to programming restrictions. This is not
286 * a restriction of the L3 implementation, but of the FF and other clients.
287 * Therefore, in a GT4 implementation it is possible for the programmed
288 * allocation of the L3 data array to provide 3*384KB=1152KB for URB, but
289 * only 1008KB of this will be used."
290 */
291 const unsigned max = (devinfo->gen == 9 ? 1008 : ~0);
292 return MIN2(max, cfg->n[GEN_L3P_URB] * get_l3_way_size(devinfo)) /
293 get_urb_size_scale(devinfo);
294 }
295
296 /**
297 * Print out the specified L3 configuration.
298 */
299 void
gen_dump_l3_config(const struct gen_l3_config * cfg,FILE * fp)300 gen_dump_l3_config(const struct gen_l3_config *cfg, FILE *fp)
301 {
302 fprintf(stderr, "SLM=%d URB=%d ALL=%d DC=%d RO=%d IS=%d C=%d T=%d\n",
303 cfg->n[GEN_L3P_SLM], cfg->n[GEN_L3P_URB], cfg->n[GEN_L3P_ALL],
304 cfg->n[GEN_L3P_DC], cfg->n[GEN_L3P_RO],
305 cfg->n[GEN_L3P_IS], cfg->n[GEN_L3P_C], cfg->n[GEN_L3P_T]);
306 }
307